Research On Underwater Distorted Image Based On Deep Learning

Due to the influence of underwater environmental factors such as water flow,underwater imaging is prone to have problems like chromatic bias,blur and distortion,which brings great difficulties to the understanding of underwater images.Therefore,the reconstruction of underwater images is extremely important,which is the basis for subsequent underwater target detection and image splitting.In recent years,thanks to the powerful automatic studying and the ability of feature extracting of deep learning,it has been inordinately used in many subjects,and fantastic progress has been made in the area of underwater image reconstruction,but on the one hand,the existing methods use shallow networks to learn about potential information,resulting in deep features cannot be fully learned.On the other hand,for large water fluctuations,the removal and image color shift restoration effect are not satisfactory.Therefore,how to build a deep network for feature extraction is the key to solving underwater image reconstruction.Based on the analysis of the research status of underwater image reconstruction at home and abroad,this paper conducts in-depth research on underwater image reconstruction in view of the above two problems,and achieves the following research results.First,we propose an underwater image reconstruction model based on U-net and image registration.In view of the problems of difficulty in finding reference pictures and image distortion in the underwater images caused by the movement of camera,we develop an indicator: Underwater Image Fluctuation Indicator(UIFI),and the index is used to evaluate the jitter degree of an underwater image so that we can choose the reference image using in the program.Furthermore,taking into account the problem of complicated structure and disappointing results of current networks,a deep residual network based on U-net architecture is designed to learn the corresponding parameters between the reference image and the floating image,and realize the depth feature registration of the image.Then,the image is reconstructed by the spatial transformation network(STN).Finally,a new normal loss function is constructed for the problems of underwater image refraction and depth distortion,and experiments indicating that the proposed algorithm can not only achieve a faster reconstruction speed than the traditional method,but also provide a new solution for underwater video understanding.Second,we put forward an underwater image reconstruction network based on feature pyramid and muti-head attention mechanism.Considering the deficiencies that the currenting methods have poor reconstruction effect on large water fluctuations and the unsatisfactory color shift recovery effect,we propose an underwater distortion image reconstruction model by using attention mechanism and Feature Pyramid.it utilizes the Feature Pyramid Network to extract deep information from all kinds of scales,so as to achieve better recovery of fluctuations,and introduces an attention mechanism to make the network learn more comprehensive color information of underwater images.Results indicate that opposed to the current method,our network performs well on different underwater datasets and obtains better reconstruction images.In summary,our methods start with an algorithm based on image registration,which uses deep learning to improve the speed of registration and accelerate the speed of reconstruction.The second image reconstruction algorithm bases on feature pyramid and multi-head attention mechanism,so that the reconstructed image does not rely on the reference image and has a wider range of application scenarios.Experimental verification shows that the proposed algorithm has better effect than the current method in underwater image reconstruction task.